In this workshop, we will finetune wav2vec2 style model from scratch on custom asr data using Fairseq library. We will also train an n-gram language model using Kenlm library. Finally we will export it to Huggingface's format and deploy it as a web app using Gradio.

Poster link here

Details and step-by-step walkthrough of training, exporting and deploying Indicwav2vec models have been outlined in the Jupyter Notebook.

Discussion on topics like ASR Pipeline, Wav2Vec2 Architecture, Role of LM in ASR, Mining Parallel Data, etc. can be found here